{"id":"0d685f2a-3a2a-4f71-bd8f-3bfb31b422e2","arxiv_id":"2506.19246","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A federated contrastive method is claimed to beat three baselines on SWaT anomaly detection, yet lacks the implementation details and error bars needed to verify the result.","lead":"This paper combines federated learning with contrastive learning to detect anomalies such as attacks and faults in distributed systems without sharing raw data. The authors report better F1, precision, and AUC than three federated baselines on the SWaT water-treatment dataset, but the paper omits code, hyperparameters, and any measure of uncertainty.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's claimed advantage over FedCAC/MOON/FedProto is unverifiable: the SWaT federated partition, windowing, train/test split, and baseline tuning are never specified, so the reported margins may reflect experimental setup rather than the method.","rationale":"The reader's weakest assumption identifies the experimental setup as the linchpin, and my independent reading reaches the same conclusion: the empirical comparison is the paper's only real contribution, and every element needed to make that comparison meaningful is missing. There is no code, no hyperparameter table, no error bars, no description of how the single SWaT time series is converted into a multi-client federated dataset, and no baseline tuning protocol. The method equations are also too malformed to permit a faithful re-implementation, which compounds the reproducibility failure. I considered whether the central concern might instead be conceptual (e.g., the supervised contrastive pairing contradicts an unsupervised detection setting), but the text explicitly mentions a classification loss, so the method is allowed to use labels; the unresolved issue is not the learning paradigm but the lack of a defined evaluation protocol. Because the claimed advantage is quantitatively small and could plausibly arise from an arbitrary partition or from undertuned baselines, the rejection is justified and no change to the reader's verdict is needed. Still, the idea of combining federated aggregation with contrastive anomaly detection has merit, so a conditional acceptance would be defensible if the authors later provide the missing protocol and reproducible code.","tokens_in":6846,"tokens_out":3275,"duration_ms":35973,"concrete_test":"Require the authors to release a runnable package, or at minimum the exact SWaT partition (sensor-to-client mapping, window length/stride, time-based train/test cut), and rerun Table 1 with a fixed evaluation harness: same data split, same thresholding rule, same optimizer budget (e.g., 100 communication rounds, 5 local epochs), and same hyperparameter search budget for all four methods over 5 seeds. If 'Ours' no longer leads on F1 across seeds, the central claim is setup-dependent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the 3.4–5.1 point F1 advantage in Table 1 (Section III.B). For that claim to hold, the comparison must be controlled: the same data assignment, windows, label usage, and tuning budget must apply to all methods. The paper supplies none of this. SWaT is a single 11-day time series; to create federated nodes the authors must partition it, but neither the number of clients nor the assignment of sensors/PLC stages to clients is given. The train/test split and sliding-window parameters are absent, as are the number of communication rounds, local epochs, batch size, temperature tau, lambda_1/lambda_2, and optimizer settings. The Section II equations are also too corrupted to pin down the architecture: the NT-Xent loss is missing its summation indices and the FedAvg update denominator is ambiguous. Because anomaly detection on SWaT can be evaluated many ways (supervised binary classification vs. unsupervised score thresholding, per-window vs. per-timestamp), the reported F1/Precision/AUC values are not anchored to a defined protocol. Under these conditions the baseline results cannot be confirmed as fairly tuned; a 3–5 point F1 gap is exactly the size that partition choice or under-tuned baselines can produce.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a method for anomaly detection in distributed systems that combines federated learning and contrastive learning. Each client encodes local behavior data into embeddings, trains with an NT-Xent contrastive loss and a classification loss, and shares model parameters with a central server for weighted aggregation. The paper evaluates the method on the SWaT dataset, reporting an F1-score of 91.5%, precision of 90.2%, and AUC of 94.7%, which it claims outperforms FedCAC, MOON, and FedProto. It also presents per-attack-type detection accuracy and a simulated real-time data stream responsiveness experiment.","tokens_in":7250,"tokens_out":4515,"duration_ms":44126,"significance":"If the reported results were reproducible, the work could be of practical interest for privacy-preserving anomaly detection in industrial systems, since it combines two established techniques in a straightforward manner. However, the paper provides no verifiable artifacts: there are no machine-checked proofs, no released code, and the experimental protocol is almost entirely unspecified. The central contribution is therefore not assessable from the manuscript as it stands, and the claimed improvements over baselines cannot be confirmed.","major_comments":[{"comment":"The central claim of a 3.4–5.1 point F1 advantage over FedCAC, MOON, and FedProto is unverifiable because the experimental protocol is not described. The manuscript does not state the number of clients, how the SWaT data are partitioned into local nodes, the sliding-window length or stride, the train/test split, the evaluation granularity (per-timestamp vs. per-window), the hyperparameters (lambda_1, lambda_2, tau, learning rate, optimizer, batch size, local epochs, communication rounds), or the tuning budget for the baselines. Without this information, the reported differences could be artefacts of the experimental setup rather than of the proposed method.","section":"Section III.B, Table 1"},{"comment":"The mathematical formulation is too corrupted to specify the method. Eq. (2) for the NT-Xent loss omits the summation indices and does not define the set Z; Eq. (3) for federated aggregation is garbled, with an unclear denominator and an ambiguous role for N, and the 'weighted average' description does not match a correct FedAvg update; Eq. (4) for the total loss is unreadable because of corrupted symbols. A reader cannot implement or verify the proposed method from these equations.","section":"Section II, Eqs. (2)–(4)"},{"comment":"The paper claims the SWaT dataset is well-suited for simulating local nodes but never specifies the actual partitioning. SWaT is a single continuous time series from a six-stage plant; the number of clients, whether partitioning is by stage, sensor group, or time range, and how heterogeneous the client distributions are, are all absent. These choices define the federated problem and materially affect the comparison.","section":"Section III.A"},{"comment":"The two additional experimental claims—per-attack accuracy and responsiveness in a simulated real-time data stream—are not supported by any provided evidence. The corresponding figures are placeholders, and the text gives only qualitative trends with approximate values (e.g., 'nearly 0.94', 'between 0.88 and 0.90'). No details are given on how the data stream is simulated, how online updates are performed, which attacks are included, or how these results are measured.","section":"Section III.B (Figures 2 and 3)"},{"comment":"No statistical support is provided for the results: the table reports single-point metrics with no standard deviations, no number of independent runs, and no significance tests. Given that the gap between methods is only 3–5 points in F1, the absence of error bars leaves open the possibility that the reported improvements are within run-to-run noise.","section":"Section III.B, Table 1"}],"minor_comments":[{"comment":"The manuscript references Figures 1, 2, and 3, but the figures are not present in the submitted text; the reader sees only placeholders.","section":"General"},{"comment":"References [13]–[20] are largely topically tangential to the claims they support (e.g., microservice scheduling, traffic scheduling, cloud memory forecasting) and do not provide meaningful methodological grounding for the contrastive loss or regularization terms.","section":"References"},{"comment":"Several equations and notations contain corruption, for example the dataset definition is rendered as '},...,,{ )()2()1( n iiii xxxD ='; these should be corrected for readability.","section":"Section II"}],"recommendation":"reject","confidential_remarks":"The reference list appears padded with self-citations and unrelated works; I recommend the editor verify the novelty and the appropriateness of the references. The manuscript is not in a publishable state due to the missing experimental protocol, corrupted equations, and absent figures."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Best to know up front: this is a standard federated-contrastive combination (MOON already pairs FedAvg with a contrastive loss) applied to SWaT, and the only evidence it works is Table 1, which lacks the experimental protocol needed to check anything. The authors may well have a working method, but this manuscript doesn't let a reader confirm it.\n\nWhat is genuinely new is narrow: adding a cross-entropy term and a local regularization penalty to the contrastive loss in a federated setting, and reporting F1/precision/AUC on SWaT. That is a reasonable configuration to try. The paper also does a sensible thing by testing on a standard ICS benchmark rather than a toy dataset, and the per-attack-type analysis and simulated stream experiment show awareness of what real deployments care about.\n\nThe soft spots are substantial. The comparison table lists three federated baselines but never says how SWaT was partitioned into clients, how windows were built, how the train/test split was done, how many communication rounds, local epochs, batch size, optimizer, or how the baselines were tuned. Without that, the 3–5 point F1 advantage is exactly the size of artifact that partition choice or under-tuned baselines can produce. The loss equations are also corrupted: the NT-Xent sum has no indices, and the FedAvg weighting formula is ambiguous. The citation pattern is odd—several references are about microservice scheduling or traffic scheduling and are used as narrative support rather than technical foundation, and one of them (ref [16]) includes an author of this paper, which is not itself a problem but doesn't add credibility here.\n\nOn the positive side, the central idea is coherent and the authors don't overclaim beyond the table. The problem is purely evidentiary: the paper gives no way to reproduce the numbers. This is a desk-reject level of reporting, not something I'd send to referees as-is. If the authors provided code, the full experimental setup, and error bars, the incremental idea might be worth a proper review. As it stands, I would not cite it and would not bring it to reading group. For peer review: reject; if a revised version with complete experimental details appears, reconsider.","headline":"Plausible incremental idea, but Table 1 is unverifiable as reported and the method section is too vague to know what was actually run.","tokens_in":7637,"tokens_out":2450,"would_cite":false,"duration_ms":24206,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that federated contrastive learning detects behavioral anomalies in distributed systems without exposing raw data, outperforming FedCAC, MOON, and FedProto on the SWaT benchmark.","keywords":["federated learning","contrastive learning","anomaly detection","distributed systems","SWaT dataset","privacy-preserving machine learning","NT-Xent loss","industrial control systems"],"falsifier":"Try to reproduce Table 1 by running the proposed method on SWaT with a fully documented client partition, fixed hyperparameters, and the same tuning budget for FedCAC, MOON, and FedProto; if the 91.5% F1 and the 3.4-point margin over MOON cannot be obtained, the central claim loses its experimental foundation.","tokens_in":6657,"feed_emoji":"🛡️","tokens_out":6371,"duration_ms":60394,"temperature":0.7,"pith_summary":"This paper proposes a method that combines federated learning with contrastive learning to detect behavioral anomalies in distributed systems while keeping raw data local. Each node learns embedding representations from logs, metrics, and system calls, then contributes to a global model through federated aggregation. On the SWaT industrial control dataset, the paper reports an F1-score of 91.5%, precision of 90.2%, and AUC of 94.7%, beating three federated baselines. If these results hold, the method offers a privacy-preserving path to accurate, real-time anomaly detection in distributed environments.","feed_headline":"Federated contrastive learning scores 91.5 F1 on SWaT anomalies","feed_subtitle":"Distributed nodes train one detector without sharing raw data, beating FedCAC, MOON, and FedProto.","key_machinery":"The carrying mechanism is the combination of a local encoder $f_\\theta$ that maps behavior data to embeddings, the NT-Xent contrastive loss that pulls same-category samples together and pushes different-category samples apart, a classification loss for predictive accuracy, and a local regularization term that keeps each node's model close to its own features while contributing to a shared model. Federated weighted averaging, $\\theta_{\\text{global}} = \\sum_i \\frac{|D_i|}{\\sum_j |D_j|} \\theta_i$, produces a global model without transmitting raw data. The contrastive loss is what the paper credits for learning the fine-grained distinctions that standard federated averaging obscures.","core_discovery":"The paper's central claim is that a federated model trained with a contrastive objective can outperform standard federated learning baselines for anomaly detection in distributed systems. Each node encodes behavior data into embeddings, forms positive and negative sample pairs from the same or different behavioral categories, and trains with NT-Xent contrastive loss, a classification loss, and a local regularization term. Parameters are aggregated by weighted averaging into a global model without raw data leaving the nodes. On SWaT, the method achieves an F1-score of 91.5%, precision of 90.2%, and AUC of 94.7%, and its accuracy rises from about 0.87 to 0.95 in a simulated real-time stream. The paper attributes these gains to contrastive separation in the embedding space and to preserving local sensitivity through regularization.","pith_inferences":["The paper leaves the federated partitioning of SWaT unspecified; a natural stress test is to vary the partition from random to physically grouped clients, and the method's stated advantage should persist if the contrastive mechanism is truly responsible.","If the same contrastive objective were applied without labels, using temporal neighbors as positive pairs and random samples as negatives, it could extend to settings where anomaly labels are scarce, a direction the paper does not explore.","The per-attack results suggest temporal contrastive learning or cross-node context fusion as a specific next step to lift DoS and timing detection above 0.88, which the paper mentions only as future work."],"forward_implications":["A global anomaly detector can be trained across nodes that never share raw behavioral data, so privacy and compliance constraints do not force centralized collection.","The reported precision of 90.2% implies fewer false alarms than the three baselines, which matters because false alarms consume diagnostic resources in distributed systems.","The attack-type breakdown shows command injection detected near 0.94 accuracy, replay and data injection between 0.88 and 0.90, and DoS and timing attacks below 0.88, indicating where future effort is needed.","In the simulated real-time stream, accuracy rises from about 0.87 to 0.95, suggesting the same architecture can serve online monitoring rather than batch analysis."],"supporting_citations":[{"why":"Supplies the federated anomaly-detection foundation for time-series data that the method extends with contrastive learning.","marker":"[12]"},{"why":"FedCAC baseline the proposed method must beat in the Table 1 comparison.","marker":"[21]"},{"why":"MOON baseline, the model-contrastive federated approach whose F1 the paper claims to exceed by 3.4 points.","marker":"[22]"},{"why":"FedProto baseline, the federated prototype-learning approach used for comparison.","marker":"[23]"}],"fun_headline_variants":["Federated contrastive learning detects anomalies without raw data","91.5 F1 on SWaT: federated contrastive detector beats baselines","Privacy-first anomaly detection via federated contrastive learning","Distributed anomaly detection: contrastive federated model hits 91.5 F1","Federated learning plus contrastive loss pinpoints subtle anomalies"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"That the comparison against FedCAC, MOON, and FedProto is fair, which requires the SWaT data partition, hyperparameters, and baseline tuning to be representative; the paper does not report this setup.","fun_headline_variants_meta":{"raw":{"variants":["Federated contrastive learning detects anomalies without raw data","91.5 F1 on SWaT: federated contrastive detector beats baselines","Privacy-first anomaly detection via federated contrastive learning","Distributed anomaly detection: contrastive federated model hits 91.5 F1","Federated learning plus contrastive loss pinpoints subtle anomalies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000263,"raw_usage":{"total_tokens":1607,"prompt_tokens":959,"completion_tokens":648,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":554}},"tokens_in":575,"tokens_out":648,"duration_ms":5685,"temperature":1.0,"reasoning_tokens":554,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:33:52.510449+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Try to reproduce Table 1 by running the proposed method on SWaT with a fully documented client partition, fixed hyperparameters, and the same tuning budget for FedCAC, MOON, and FedProto; if the 91.5% F1 and the 3.4-point margin over MOON cannot be obtained, the central claim loses its experimental foundation.","supporting_citations":[{"cited_title":"Light-weight federated learning-based anomaly detection for time-series data in industrial control systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the federated anomaly-detection foundation for time-series data that the method extends with contrastive learning."},{"cited_title":"Bold but cautious: Unlocking the potential of personalized federated learning through cautiously aggressive collaboration,","cited_arxiv_id":null,"evidence_quote":"FedCAC baseline the proposed method must beat in the Table 1 comparison."},{"cited_title":"Model-contrastive federated learning,","cited_arxiv_id":null,"evidence_quote":"MOON baseline, the model-contrastive federated approach whose F1 the paper claims to exceed by 3.4 points."},{"cited_title":"Fedproto: Federated prototype learning across heterogeneous clients,","cited_arxiv_id":null,"evidence_quote":"FedProto baseline, the federated prototype-learning approach used for comparison."}],"review_version":1}